Trellis 2 - 3D Model Gen from Microsoft on DGX Spark
- URL: https://forums.developer.nvidia.com/t/trellis-2-3d-mode-gen-from-microsoft/355030/15
- Date Saved: 2026-06-23
- Source: NVIDIA Developer Forums (DGX Spark / GB10 Projects)
- Tags: creative-ai, local-llm
Summary
Forum thread about running Microsoft’s TRELLIS.2 (3D model generation from HuggingFace) on NVIDIA DGX Spark.
Key challenges encountered:
- externally-managed-environment error — Docker builds fail because newer Debian/Ubuntu Python refuses pip installs without
--break-system-packagesor a venv - pillow-simd won’t build on aarch64 (Spark is ARM-based GB10)
- CUDA toolkit 12.x unavailable for the platform, needed for flash-attention
- PTX errors when trying CUDA 13.1 as fallback
- General theme: nothing works “straight up” on Spark due to non-standard ARM architecture + missing prebuilt wheels
Solution (post #19, by propellerheadvij):
- Created a working Docker setup: https://github.com/dr-vij/Trellis2-DGX-Spark-Docker
- Modified Gradio app.py to allow configuring IP/port (for macOS remote access)
- Confirmed working — “generation quality is quite decent”
- Got 5 upvotes, marked as accepted solution by another user (iftgaf)
Post #15 (linked specifically):
User deeduckme hit the PEP 668 “externally-managed-environment” error during Docker build step 3/15 when running pip install --upgrade pip setuptools wheel. Fix: add --break-system-packages flag or use a venv inside the Dockerfile.
Related threads:
- “TRELLIS.2 on DGX Spark” (separate project post by RazielAU with a fork)
- Flash Attention 3 install failures on Spark (common theme)
- vLLM Docker images for DGX Spark (156 replies, very active)